Lecture 4

Always plot your data!

All of these data sets have the same average and standard deviation for their x and y variables, and the same correlation between x and y.

Visualizing the distribution of a continuous variable

A histogram is a common and useful way to visualize the distribution of a continuous variable. A histogram is a special bar chart where the categories along the x-axis are bins (i.e. ranges) of values for a variable, and the heigth of the bars indicates the number of observations within each bin.

If you have a data set loaded into R as a data frame, you can quickly create a histogram using the geom_histogram() function within ggplot(). In this example, I have a data frame called crashes and one of the variables in that data frame is called crashes_per_km2. The bins argument indicates the number of bins I want in my histogram.

ggplot(crashes) +
  geom_histogram(aes(x = crashes_per_km2),
                 color = "black",
                 fill = "pink",
                 bins = 30) +
  theme_minimal()

You can also create a histogram in Excel. You need to define a set of threshold values for your bins, then use the “Histogram” tool in the “Data Analysis” toolpack, which you’ll find on the “Data” ribbon if you’ve enabled it.

A jittered (or one-dimensional) scatter plot can be another good way to visualize the distribution of a variable. Making your points very small (size = 0.5) and transparent (alpha = 0.5) can be helpful if you have a lot of points plotted on top of one another (a problem called overplotting).

ggplot(crashes) +
  geom_point(aes(x = crashes_per_km2, y = 0),
             position = "jitter",
             size = 0.5,
             alpha = 0.5) +
  scale_y_continuous(name = "",
                     breaks = c()) +
  theme_minimal() 

Visualizing the relationship between two continuous variables

A more typical use of a scatter plot is to visualize the relationship between two continuous variables. In this example, I want to see the relationship between population density (people per square kilometer) and crash density (crashes per square kilometer). It is standard to have your outcome variable on the y-axis.

ggplot(crashes) +
  geom_point(aes(x = ppl_per_km2, 
                 y = crashes_per_km2),
             size = 0.5,
             alpha = 0.5) +
  theme_minimal() 

It’s straightforward to create a plot like this in Excel as well.

By default, Excel will guess what values you want on your x and y axes, and it sometimes guesses wrong. I recommend clearing out all the data series and adding in the one you want manually to be sure that you’re plotting what you think you’re plotting.

Decribing continuous variables

You can describe a continuous variable in terms of its central tendency and its spread.

mean(crashes$crashes_per_km2)

## [1] 368.2335

quantile(crashes$crashes_per_km2, probs = c(0.25, 0.5, 0.75))

##      25%      50%      75% 
## 154.8500 324.3568 519.0817

summary(crashes$crashes_per_km2)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   34.93  154.85  324.36  368.23  519.08 1906.71

Describing the distribution of a continuous variable

In real life, it’s common for variables to fall into histograms that are shaped kind of like a bell (a bell curve). This is so common, that we call it a normal distribution. It looks something like this, with more values closer the the average value, and fewer values farther from the average. It’s symmetrical in the sense that values are as likely to be below average as above average (so the median will be almost the same as the mean).

The methods we discuss in this class are based on an assumption that your continuous variables have an approximately normal distribution. Sometimes they don’t. This one doesn’t.

When you have a lot of lower values and fewer higher values like this, you can call this a left-skewed distribution (because there are more values on the left side of the histogram than on the right). A left-skewed distribution might indicate a log-normal distribution, where the logarithms of the values approximate a normal distribution better than the values themselves do.

You might also have a bi-modal distribution like the one below. Where the histogram shows two distinct humps. This might suggest you have two different populations (or categories) in your data, and you might need to account for that in your analysis somehow.

Summarizing a categorical variable

Variation in a categorical variable is somewhat simpler. You’ve more or less described it fully when you’ve reported the proportion of your sample in each category.

If I have a data frame in R called commuting, with a variable called mode to indicate the typical mode of transportation to work, I can get the number of people who report using each mode like this:

commuting %>%
  group_by(mode) %>%
  summarise(number = n())

## # A tibble: 8 × 2
##   mode       number
##   <chr>       <int>
## 1 Bike         8390
## 2 Car        715885
## 3 Ferry         656
## 4 Motorcycle   3011
## 5 Other        8555
## 6 Taxi         1460
## 7 Transit     42950
## 8 Walk        23893

And I can show that as the proportions by dividing by the total sample size (which would be the length of any of the data frame’s columns).

commuting %>%
  group_by(mode) %>%
  summarise(number = n()) %>%
  mutate(proportion = number / length(commuting$mode)) 

## # A tibble: 8 × 3
##   mode       number proportion
##   <chr>       <int>      <dbl>
## 1 Bike         8390   0.0104  
## 2 Car        715885   0.890   
## 3 Ferry         656   0.000815
## 4 Motorcycle   3011   0.00374 
## 5 Other        8555   0.0106  
## 6 Taxi         1460   0.00181 
## 7 Transit     42950   0.0534  
## 8 Walk        23893   0.0297

I can do something similar in Excel using a pivot table:

Visualizing proportions

commuting %>%
  group_by(mode) %>%
  summarise(number = n()) %>%
  mutate(proportion = number / length(commuting$mode)) %>%
  
  ggplot() +
    geom_bar(aes(x = mode, y = proportion),
             stat = "identity") +
    scale_y_continuous(breaks = breaks <- seq(0, 1, by = 0.1),
                       labels = paste0(breaks * 100, "%")) +
    theme_minimal()  

library(treemapify)

commuting %>%
  group_by(mode) %>%
  summarise(number = n()) %>%
  mutate(proportion = number / length(commuting$mode)) %>%
  
  ggplot() +
    geom_treemap(aes(area = proportion, fill = mode)) +
    geom_treemap_text(aes(area = proportion, label = mode),
                      color = "white") +
    scale_fill_brewer(type = "qual", 
                      guide = "none") # leave off the legend